Tactile Data for Robots: Five Approaches, One Unanswered Cost Question
New tactile robot datasets and models expose diverging costs for scaling contact data, alongside gains in real-time control and evaluation rigor.
OpenNeoData, released by NeoteAI and Fudan, is the largest tactile-robotics dataset yet: 200,000+ trajectories and 5,000+ hours across six embodiments, with gel visuotactile cameras time-aligned to RGB at 30fps and shipped in LeRobot v3.0 format. Two accompanying models, N₀-VTLA and N₀-TWAM, claim to be the first VLA pretrained at scale on tactile data and a world-action model predicting future contact alongside vision, respectively.
Four other teams sidestep the cost of real sensors entirely. Being-H0.8 infers contact fields from 500,000 hours of ordinary human video via a model called TactoHand; ViTacWorld leans on simulation, arguing tactile signal crosses the sim-to-real gap better than vision; τ learns a tactile representation only during training, adding nothing at inference; FeelWorld models contact, force, and slip as separate gated states. Despite different price tags, four of five converge on predicting future tactile state as the core objective — the real disagreement is over what an hour of contact data should cost, and only OpenNeoData's real measurements are actually downloadable, gated behind a contact form.
Elsewhere in robot learning: FLUX-mimic repurposes video-prediction features for real-time action decoding, reporting 95% success on a soft-body kitting task versus 55% for a baseline; HiFi-UMI shows that high-fidelity teleoperation capture can match in-domain performance without needing a real-robot anchor; πR² splits fast and slow conditioning channels to replan roughly 4x faster; and the Curse of Precision paper quantifies how demonstration counts scale super-exponentially as target precision tightens. AXIS and ArmnetBench tackle data-collection and evaluation infrastructure at scale, while MiniCPM-RobotManip open-sources a compact 1.5B generalist manipulation model under Apache-2.0.